Demis Hassabis, founder of Google DeepMind and Nobel laureate in chemistry, now believes artificial general intelligence (AGI) could arrive within two to three years. It is a further shortening: shortly before he spoke of three to five years, and in January of a ten-year horizon. "It's going to happen soon. My estimate keeps moving closer to the present," he told WSJ Magazine, which on October 21 published a major profile of him.
According to Hassabis, AGI is a system matching the full cognitive range of the human brain. To test it he uses what he calls the "Einstein test": could a system trained only on pre-1901 knowledge reproduce Einstein's 1905 papers on its own? The idea, reported by Zaruko, is that a true AGI should get there independently, as the physicist did.
Image: Wikipedia — https://en.wikipedia.org/wiki/Demis_Hassabis
In the profile Hassabis compares AGI's impact to "ten industrial revolutions at ten times the speed." Since stepping back from day-to-day lab management in August, he has focused mainly on drug discovery at Google DeepMind, where his forecast is more concrete. On safety he is more cautious: the same week, in Dublin, he warned that safety infrastructure is not ready for what is coming, and proposed an industry-funded oversight body under government control to vet models before release. Bad regulation, he argues, is worse than none.
An earlier warning
The forecast is not an isolated statement: in September Hassabis had already warned that AI systems could cause serious safety incidents within two years, calling for international standards and sandbox testing. Oversight is also part of DeepMind's strategy, which has previously discussed an independent FINRA-style body to test frontier models before release.
The timing remains contested. The debate over when AGI will arrive has shifted to data: some point to real progress stuck at around 65% of the AI 2027 forecasts, while others, like Anthropic's Jack Clark, have brought recursive self-improvement forward to 2027. Sam Altman, meanwhile, is aiming for an internal AGI system by the end of 2026, though development of the Astra model is paused over misalignment.
No comments yet. Be the first!